Baretools AI

Why Baretools?

← Home  ·  Get Started  ·  Concepts  ·  Providers  ·  API Reference

---

The Hidden Cost of Heavy Frameworks

When you pip install langchain, you’re not installing one package. You’re pulling in a dependency tree that includes 50–100+ transitive packages: Pydantic, httpx, tenacity, SQLAlchemy, aiohttp, tiktoken, numpy, and more — many of which you’ll never use.

Each of those packages is:

• A potential source of CVEs you need to track and patch
• A package whose maintainer could push a compromised release (see: event-stream, xz-utils, polyfill.io)
• A version constraint that blocks your other deps from upgrading
• Megabytes of code that slows container builds and expands your attack surface

Baretools’ Approach

$ pip install baretools-ai
$ pip show baretools-ai | grep Requires
Requires:

Nothing. Baretools uses only the Python standard library:

Feature	Standard library module
Type hints and schema reflection	inspect, typing
Async execution	asyncio
Parallel execution	concurrent.futures
Structured types	dataclasses, typing.TypedDict
Logging / events	logging
JSON parsing	json

Dependency Comparison

The table below compares runtime dependency counts for popular tool-calling solutions. Figures are approximate as of April 2026 — install in a fresh venv with pip list to verify.

Library	Runtime deps (approx.)	Install size (approx.)	Notes
baretools-ai	0	~25 KB	stdlib only
langchain-core	~15	~5 MB	pydantic, httpx, tenacity, yaml…
langchain + langchain-openai	~50	~30 MB	adds tiktoken, openai SDK, etc.
llama-index-core	~30	~15 MB	pydantic, httpx, nltk, numpy…
crewai	~40+	~20 MB	langchain, pydantic, litellm…

When the bigger frameworks are the right call. If time-to-market is the dominant constraint — you need agents, retrieval, memory, prompt templates, and evals stitched together this week — LangChain, LlamaIndex, and CrewAI will get you there faster than rolling your own. The trade-off you’re accepting is a larger supply chain to audit, less control over the agent loop, and a steeper learning curve for the framework itself versus the underlying APIs. Baretools is for the case where that trade-off has stopped paying off.

What We Deliberately Exclude

Pydantic — a great library, but not everyone needs it. If your tools accept BaseModel parameters, install it explicitly: pip install "baretools-ai[pydantic]". Otherwise, standard-library @dataclass types work out-of-the-box with zero extra installs.

httpx / requests — Baretools never makes network calls. HTTP is your application’s responsibility.

tiktoken / tokenizers — Token counting and context management are orchestration concerns that vary by application. They belong in your code.

LangChain / LangGraph — Baretools is not an alternative to the orchestration parts of those libraries. It replaces only the tool-wiring plumbing, which you can now handle yourself without pulling in the rest.

Supply-Chain Risk in Practice

The Python Package Index has seen a sustained rise in typosquatting, dependency confusion, and maintainer account compromise attacks. The fewer packages in your dependency tree, the smaller your blast radius.

A zero-dependency library means:

• No upstream compromise can reach your tool execution layer via a transitive package
• pip-audit and safety scans complete instantly — there is nothing to scan
• Renovate / Dependabot produce zero noise for the baretools install itself
• Security reviews are fast: read src/baretools/core.py (~900 lines of pure Python) and you’re done

Philosophy

Baretools was designed around one constraint: if a feature requires a non-stdlib dependency, it does not go in core.

The corollary is that Baretools is intentionally small. It will never include:

• An LLM client (use the provider’s official SDK)
• Prompt templates (use f-strings or your own engine)
• Agent orchestration (your loop, your rules)
• Vector stores, memory, or retrieval

If you need those things, excellent libraries exist for each. Install only what you need.

This site is open source. Improve this page.